
Work Here?
Liquid AI builds and deploys foundation models based on liquid neural networks, focusing on efficient, on-device AI. It develops Liquid Foundation Models (LFMs), a family of generative AI models designed to be smaller and more computation-efficient than typical large language models, enabling deployment on edge devices with lower latency, better privacy, and reduced infrastructure costs. The approach includes end-to-end AI expertise and customizable architectures for enterprises that require real-time performance and private processing. Compared to traditional AI providers, Liquid AI emphasizes edge-ready, adaptable models that run efficiently on constrained hardware, and it targets enterprise-grade, private, reliable AI solutions. The company’s goal is to enable real-time, on-device AI at scale for businesses by offering compact, capable foundation models and the tools to customize them for specific applications.
Industries
Data & Analytics
Enterprise Software
AI & Machine Learning
Company Size
51-200
Company Stage
Series A
Total Funding
$287.5M
Headquarters
Brookline, Massachusetts
Founded
2023
People at Liquid AI who can refer or advise you
Help us improve and share your feedback! Did you find this helpful?
Total Funding
$287.5M
Above
Industry Average
Funded Over
2 Rounds
Industry standards
Remote Work Options
Flexible Work Hours
MIT in the media: For the future of tech, "Massachusetts can absolutely lead" MIT News. On June 9, The Boston Globe released its 2026 "Tech Power Players" list, recognizing 50 influential local leaders in technology and business across Massachusetts. The list includes eight MIT affiliates including President Sally Kornbluth, Prof. Daniela Rus (director of CSAIL), Prof. Regina Barzilay, Prof. Yet-Ming Chiang, Prof. Max Tegmark, Ana Bakshi (executive director of the Martin Trust Center for MIT Entrepreneurship), Katie Rae CEO and Managing Partner of Engine Ventures), and Senior Lecturer Brian Halligan, along with a number of MIT alumni. In addition to recognizing individual leaders, the Power Players coverage highlights MIT's research labs, its culture of innovation and entrepreneurship, industry connections, new AI initiatives, and the Institute's deep commitment to maintaining Massachusetts' technological leadership. "Massachusetts can absolutely lead in this next wave," says President Kornbluth, noting that the future is bright with burgeoning opportunities to advance technologies in fields from manufacturing, life and health sciences to quantum technologies and energy in service of Americans across the country. Advancing AI and entrepreneurship When it comes to AI, MIT is "working to drive artificial intelligence forward in sectors where the region is strongest, from biotechnology and robotics to defense and clean energy. It's also trying to broaden entrepreneurship through a 'dorm-to-startup' push, creating a pipeline of support services - from hack-a-thons to venture funding - to help students to start companies between classes," writes Robert Weisman for The Globe. Looking ahead, The Globe highlights how MIT aims to remain a central driver of AI advancement within higher ed. "President Sally Kornbluth is reinvigorating the school's support of the local innovation ecosystem," writes Aaron Pressman, noting how MIT is "unveiling new online classes dedicated to AI - with free entry-level classes for anyone - and encouraging more entrepreneurship on campus." MIT's free, online AI courses could help local tech leaders in their challenge "to ensure people, not only corporations, benefit from the technology," writes Pressman. And when it comes to applying AI technologies to real-world problems, MIT aims to ensure the greater Boston area remains a leader. "Some schools in Massachusetts, including MIT, are carving out a specialty in applied AI - sometimes called 'AI+X' - deploying the technology to help businesses, hospitals, and research institutions to supercharge productivity, innovation, and scientific breakthroughs," explains Weisman. Aman Narang '04, CEO of Toast, adds: "The superpower has always been the university system. The best thing Boston can do is keep these people around." MIT startups are a key driver of the region's entrepreneurial ecosystem. To ensure the greater Boston area remains a hub for innovators and to respond to growing student interest, MIT is looking to build upon its existing entrepreneurship resources for students, including the more than 150 courses and 85 centers and programs dedicated to fostering an entrepreneurial community. Additionally, President Sally Kornbluth and Provost Anantha Chandrakasan recently formed the Committee on Accelerating Translation and Entrepreneurship (CATE) to explore anew how the Institute can best support, remove barriers to, and accelerate the movement of ideas from MIT's research and innovative discoveries into new ventures. Further, reflecting on the optimism surrounding the Greater Boston tech scene, The Globe describes how applications for The Martin Trust Center for MIT Entrepreneurship's startup accelerator program have doubled from last year, and nearly one-fifth of MIT undergraduates - about 800 students - attended a recent startup career fair. Innovating change beyond MIT The simple worm could drive the future of AI. This might sound like a squishy premise, but that's the idea behind MIT startup Liquid AI, which is developing AI models inspired by the brain structure of a simple worm and could significantly reduce AI energy consumption. Liquid AI's models, "which can uncover financial fraud and pilot autonomous drones, require far less electricity to operate than large language models, saving energy and water, which is used to cool data centers," Pressman explains. The Globe highlights how Liquid AI recently signed a deal with Mercedes-Benz to incorporate its technology into the onboard systems of cars sold in North America. To power new AI technologies - and ensure Americans across the country can have reliable and affordable energy sources - researchers at MIT and a number of alumni are also turning their attention to the future of energy. In Prof. Yet-Ming Chiang's lab, researchers are developing batteries that can store more electricity over longer periods, creating "more opportunities for wind, solar, and other clean energy sources." Weisman highlights how "Chiang's lab and other MIT research centers are also working on innovations in microchips, critical minerals, fusion technology, and defense tech. All are examples of 'tough tech' projects combining science and engineering, which Chiang says 'are in the sweet spot of the Boston ecosystem.'" Soon, 80 MIT students will work as summer interns and employees at GE Vernova, thanks to the MIT-GE Vernova Climate and Energy Alliance, a collaboration aimed at advancing research and education that will accelerate the global energy transition. GE Vernova CEO Scott Strazik wanted his organization to "plug into the city's innovation culture," particularly the MIT campus and community. The company announced it would dedicate $50 million over five years to fund internships and research projects in which students and faculty work alongside GE Vernova engineers and technicians. The most promising area for the Greater Boston tech scene The Globe concludes by asking each Power Player what the most promising thing about the Greater Boston tech scene is right now. For Rus, the answer is: "talent. Boston has the best AI researchers in the world, and they're producing genuinely new ideas, not incremental ones," she explains. When it comes to realizing the potential of fusion energy, Bob Mumgaard SM '15, co-founder and CEO of Commonwealth Fusion Systems, explains that he couldn't have built the company anywhere but Massachusetts thanks to the region's expertise in engineering, designing, and manufacturing hardware and equipment and access to university researchers. "The ecosystem has the building blocks," says Mumgaard. "Massachusetts is the strongest in the nation in innovation in energy." President Kornbluth points to quantum. "There isn't a more important technological field right now than quantum science and technology, and the Boston area has the greatest concentration of quantum talent anywhere in the world," Kornbluth emphasizes. Loading...
Insilico Medicine releases 2025 annual results and advances AI drug discovery platform. March 29, 2026 at 4:15 PM - by MLQ Agent Key points. * Insilico Medicine will report 2025 financial results and business update on March 30, 202613. * Live conference calls in English and Mandarin set for 9:00 AM and 10:30 AM Beijing Time13. * Company advancing AI drug discovery with recent collaborations and platform enhancements57. * Past quarters show revenue declines averaging 21.7% annually alongside R&D investments6. * Recognized for AI platform impact, including Phase IIa results for rentosertib2. Insilico Medicine, a clinical-stage biotech firm using generative AI for drug discovery, plans to release its 2025 financial results and business update on March 30, 2026. The announcement follows a March 13 press release detailing conference calls in English and Mandarin13. Announcement details. Insilico Medicine announced on March 13, 2026, that it will report financial results for the year ended December 31, 2025, during live conference calls on March 30, 2026, Beijing Time. The English session starts at 9:00 AM Beijing Time, equivalent to 9:00 PM U.S. Eastern Time on March 29, 2026. The Mandarin session follows at 10:30 AM Beijing Time. Participants must pre-register via the provided Zoom link for the English call 134. Replays will be available on the company's website shortly after the calls 1. Recent business highlights. Insilico recently collaborated with Liquid AI, announced on March 8, 2026, focusing on AI advancements for drug discovery 57. In 2025, the company reported Phase IIa results for rentosertib, its first fully AI-discovered and designed small-molecule drug, which showed lung function stabilization or improvement in idiopathic pulmonary fibrosis patients with a favorable safety profile 2. The firm also completed Hong Kong's largest biotech IPO in 2025 and expanded partnerships with global pharmaceutical companies 2. Financial background. Historical data indicates revenues declining at an average annual rate of 21.7%, with net margins at -82.72% as of the last update 6. Quarterly figures for 2025 show revenues of $54 million in Q2 with a $44 million loss, following $70 million in Q1 with a $31 million loss. R&D expenses remained high, at $82 million in Q2 and $87 million in Q1 6. AI platform investment returns. Insilico Medicine's persistent revenue decline of 21.7% annually contrasts with robust R&D spending, reflecting heavy investment in its generative AI platform amid a challenging biotech funding environment 6. The company's recognition by Fast Company for clinical progress, particularly rentosertib's Phase IIa data, underscores the platform's potential to deliver real-world outcomes, even as losses widened 2. This pattern aligns with clinical-stage biotechs prioritizing pipeline advancement over near-term profitability. Strategic collaborations like the March 2026 Liquid AI milestone on private infrastructure highlight Insilico's focus on enhancing AI capabilities for drug discovery applications 57. The upcoming earnings call timing allows investors to assess how these developments offset revenue pressures and one-off costs, positioning the firm amid growing competition in AI-biotech integration. Earnings call pipeline updates. Investors await March 30 conference calls for insights into 2025 performance, including potential updates on rentosertib progression and new pipeline assets 13. Replays and website access will broaden reach, potentially influencing stock movements for ticker 3696.HK. Expanding partnerships could accelerate revenue through milestones or licensing 2 . Longer-term, Insilico's end-to-end AI approach may drive efficiencies in drug development timelines, building on 2025's clinical milestones. Sustained R&D amid revenue challenges will test capital allocation post-IPO, with focus on advancing preclinical and clinical programs toward commercialization 62. Market reception to earnings guidance will signal confidence in AI-driven biotech scalability. Further sources. Written with AI assistance, verified and edited by its team. Questions? Contact MLQ.ai.
Liquid AI releases LocalCowork powered by LFM2-24B-A2B to execute Privacy-First agent workflows locally via Model Context Protocol (MCP). March 5, 2026 Liquid AI has released LFM2-24B-A2B, a model optimized for local, low-latency tool dispatch, alongside LocalCowork, an open-source desktop agent application available in their Liquid4All GitHub Cookbook. The release provides a deployable architecture for running enterprise workflows entirely on-device, eliminating API calls and data egress for privacy-sensitive environments. Architecture and Serving configuration. To achieve low-latency execution on consumer hardware, LFM2-24B-A2B utilizes a Sparse Mixture-of-Experts (MoE) architecture. While the model contains 24 billion parameters in total, it only activates approximately 2 billion parameters per token during inference. This structural design allows the model to maintain a broad knowledge base while significantly reducing the computational overhead required for each generation step. Liquid AI stress-tested the model using the following hardware and software stack: * Hardware: Apple M4 Max, 36 GB unified memory, 32 GPU cores. * Serving Engine: llama-server with flash attention enabled. * Quantization: Q4_K_M GGUF format. * Memory Footprint: ~14.5 GB of RAM. * Hyperparameters: Temperature set to 0.1, top_p to 0.1, and max_tokens to 512 (optimized for deterministic, strict outputs). LocalCowork Tool Integration. LocalCowork is a completely offline desktop AI agent that utilizes the Model Context Protocol (MCP) to execute pre-built tools without relying on cloud APIs or compromising data privacy, logging every action to a local audit trail. The system includes 75 tools across 14 MCP servers capable of handling tasks like filesystem operations, OCR, and security scanning. However, the provided demo focuses on a highly reliable, curated subset of 20 tools across 6 servers, each rigorously tested to achieve over 80% single-step accuracy and verified multi-step chain participation. LocalCowork acts as the practical implementation of this model. It operates completely offline and comes pre-configured with a suite of enterprise-grade tools: * File Operations: Listing, reading, and searching across the host filesystem. * Security Scanning: Identifying leaked API keys and personal identifiable information (PII) within local directories. * Document Processing: Executing Optical Character Recognition (OCR), parsing text, diffing contracts, and generating PDFs. * Audit Logging: Recording every tool call locally for compliance tracking. Performance benchmarks. Liquid AI team evaluated the model against a workload of 100 single-step tool selection prompts and 50 multi-step chains (requiring 3 to 6 discrete tool executions, such as searching a folder, running OCR, parsing data, deduplicating, and exporting). Latency. The model averaged ~385 ms per tool-selection response. This sub-second dispatch time is highly suitable for interactive, human-in-the-loop applications where immediate feedback is necessary. Accuracy. * Single-Step Executions: 80% accuracy. * Multi-Step Chains: 26% end-to-end completion rate. Key takeaways. * Privacy-First Local Execution: LocalCowork operates entirely on-device without cloud API dependencies or data egress, making it highly suitable for regulated enterprise environments requiring strict data privacy. * Efficient MoE Architecture: LFM2-24B-A2B utilizes a Sparse Mixture-of-Experts (MoE) design, activating only ~2 billion of its 24 billion parameters per token, allowing it to fit comfortably within a ~14.5 GB RAM footprint using Q4_K_M GGUF quantization. * Sub-Second Latency on Consumer Hardware: When benchmarked on an Apple M4 Max laptop, the model achieves an average latency of ~385 ms for tool-selection dispatch, enabling highly interactive, real-time workflows. * Standardized MCP Tool Integration: The agent leverages the Model Context Protocol (MCP) to seamlessly connect with local tools - including filesystem operations, OCR, and security scanning - while automatically logging all actions to a local audit trail. * Strong Single-Step Accuracy with Multi-Step Limits: The model achieves 80% accuracy on single-step tool execution but drops to a 26% success rate on multi-step chains due to 'sibling confusion' (selecting a similar but incorrect tool), indicating it currently functions best in a guided, human-in-the-loop loop rather than as a fully autonomous agent.
Insilico Medicine and Liquid AI have partnered to create LFM2-2.6B-MMAI, a lightweight scientific foundation model for drug discovery that runs entirely on private pharmaceutical infrastructure. The 2.6-billion-parameter model achieves state-of-the-art performance across multiple drug discovery tasks whilst being ten times smaller than comparable systems. The model covers property prediction, molecular optimisation, affinity prediction and chemical reasoning. It outperformed TxGemma-27B on 13 of 22 pharmacokinetics and toxicology tasks, achieved 98.8% success rates on multi-parameter optimisation benchmarks, and produced better correlation scores than GPT-5.1, Claude Opus 4.5 and Grok-4.1 on Insilico's internal benchmark featuring 2.5 million experimental measurements across 689 protein targets. The collaboration addresses pharmaceutical companies' need to use advanced AI capabilities without sending proprietary data to external cloud services.
Liquid AI releases LFM2.5-1.2B-Thinking: a 1.2B parameter reasoning model that fits under 1 GB on-device. Liquid AI has released LFM2.5-1.2B-Thinking, a 1.2 billion parameter reasoning model that runs fully on device and fits in about 900 MB on a modern phone. What needed a data center 2 years ago can now run offline on consumer hardware, with a focus on structured reasoning traces, tool use, and math, rather than general chat. Position in the LFM2.5 family and core specs. LFM2.5-1.2B-Thinking is part of the LFM2.5 family of Liquid Foundation Models, which extends the earlier LFM2 architecture with more pre-training and multi stage reinforcement learning for edge deployment. The model is text only and general purpose with the following configuration: * 1.17B parameters, reported as a 1.2B class model * 16 layers, with 10 double gated LIV convolution blocks and 6 GQA blocks * Training budget of 28T tokens * Context length of 32,768 tokens * Vocabulary size of 65,536 * 8 languages, English, Arabic, Chinese, French, German, Japanese, Korean, Spanish Reasoning first behavior and thinking traces. The 'Thinking' variant is trained specifically for reasoning. At inference time it produces internal thinking traces before the final answer. These traces are chains of intermediate steps that the model uses to plan tool calls, verify partial results, and work through multi step instructions. Liquid AI team recommends this model for agentic tasks, data extraction pipelines, and retrieval augmented generation flows where you want explicit reasoning and verifiable intermediate steps. A practical way to think about it, you use LFM2.5-1.2B-Thinking as the planning brain inside agents and tools, and use other models when you need broad world knowledge or code heavy workflows. Benchmarks versus other 1B class models. Liquid AI team evaluates LFM2.5-1.2B-Thinking against models around 1B parameters on a suite of reasoning and instruction benchmarks. Compared to LFM2.5-1.2B-Instruct, three metrics improve strongly, math reasoning rises from about 63 to 88 on MATH 500, instruction following rises from about 61 to 69 on Multi IF, and tool use rises from about 49 to 57 on BFCLv3. LFM2.5-1.2B-Thinking competes with Qwen3-1.7B in thinking mode on most reasoning benchmarks while using around 40 percent fewer parameters and fewer output tokens on average. It also outperforms other 1B class baselines such as Granite-4.0-H-1B, Granite-4.0-1B, Gemma-3-1B-IT, and Llama-3.2-1B Instruct on many of these tasks. Training recipe and doom looping mitigation. Reasoning models often suffer from doom looping, where the model repeats fragments of its chain of thought instead of finishing the answer. LFM2.5-1.2B-Thinking uses a multi stage training pipeline to reduce this. The process starts with mid training that includes reasoning traces so the model learns a 'reason first then answer' pattern. Then supervised fine tuning on synthetic chains improves chain of thought generation. After that, preference alignment and RLVR are applied. In preference alignment, the research team generates 5 temperature sampled candidates and 1 greedy candidate per prompt and uses an LLM judge to pick preferred and rejected outputs, while also labeling looping outputs explicitly. During RLVR they add an n gram repetition penalty early in training. This reduces the doom loop rate from 15.74 percent at mid training to 0.36 percent after RLVR on a set of representative prompts. The result is a small reasoning model that can produce thinking traces without getting stuck in long repetitive outputs, which is important for interactive agents and on device UX. Inference performance and hardware footprint. A key design target is fast inference with a small memory footprint on CPUs and NPUs. LFM2.5-1.2B-Thinking can decode at about 239 tokens per second on an AMD CPU and about 82 tokens per second on a mobile NPU, while running under 1 GB of memory, with broad day one support for llama.cpp, MLX, and vLLM. The detailed hardware table uses 1K prefill and 100 decode tokens and gives the following examples for LFM2.5-1.2B-Thinking These numbers show that the model fits comfortably under 1 GB on phones and embedded devices while sustaining useful throughputs even at long contexts. Key takeaways. * LFM2.5-1.2B-Thinking is a 1.17B parameter reasoning model with 32,768 context length and runs under 1 GB on phones and laptops. * The model is optimized for explicit thinking traces, agentic workflows, data extraction, and RAG. * It reaches strong scores for a 1B class model, for example 87.96 on MATH 500, 85.60 on GSM8K, and competitive performance with Qwen3 1.7B in thinking mode with fewer parameters. * The training pipeline uses midtraining with reasoning traces, supervised fine tuning, preference alignment with 5 sampled along with 1 greedy candidate, and RLVR with n gram penalties, which reduces doom loops from 15.74 percent to 0.36 percent. * The model runs efficiently on AMD and Qualcomm NPUs and CPUs with runtimes like llama.cpp, FastFlowLM, and NexaML, is available in GGUF, ONNX, and MLX formats, and can be loaded easily from Hugging Face for on device deployment. Hosting providers/deployment. You can access or host the model through the following providers and platforms: Cloud & API providers. Model repositories (self-hosting). If you want to run the model locally or on your own infrastructure, the weights are available in various formats:
Find jobs on Simplify and start your career today
Industries
Data & Analytics
Enterprise Software
AI & Machine Learning
Company Size
51-200
Company Stage
Series A
Total Funding
$287.5M
Headquarters
Brookline, Massachusetts
Founded
2023
Find jobs on Simplify and start your career today